In this experiment, we use Ego-VQA dataset. Please cite this paper[27] if you use this dataset. The original IU Multi-view egocentric video dataset can be downloaded [here]. Though it's not necessary to download the original frames, it would be helpful to visualize the egocentric data.
Because the Ego-VQA is a generally small QA dataset, we use a large YouTube2Text-QA dataset for pre-training. Please cite this paper[27] if you use this dataset.
python pretrain.py --memory_type=_mrm2s | _stvqa | _enc_dec | _co_mem
Please manually set select the best model on validation set for each model (memory_type), and modify train.py accordingly to update the initialization models.
For training and validating, execute the following command
python train.py --split=1|2|3 --memory_type=_mrm2s | _stvqa | _enc_dec | _co_mem
For testing, execute the following command
python test.py --memory_type=_mrm2s | _stvqa | _enc_dec | _co_mem
Please modify train.py to manually add your pre-trained models to initialize the entire model. Current we use default models we provided in previous steps.
Python = 2.7
PyTorch = 1.0+ [here]
GPU training with 4G+ memory, testing with 1G+ memory.